Recent CNN Advancements For Stratification of Hyperspectral Images

Abstract

Stratification of hyperspectral images has become an essential in the area of remote sensing having the capability to analyze and categorize diversified land cover. Several classification models for hyperspectral images have been proposed. On one hand, conventional machine learning techniques struggled to retrieve discriminative features from HSI because of strongly correlated bands and scarcity of limited data. However recently introduced deep learning methods have recently been acknowledged as effective extraction of features techniques, having the capability to show great classification performance even with limited training data. Convolutional neural networks (CNNs) in specific are extremely efficient and have the potential to produce high performance in HSI classification. Inspired by the overall success of CNNs, this paper thoroughly examines state-of-the-art CNN architectures involved in classifying hyperspectral images. We focus on current convolutional networks that can retrieve spectral or spatial or spectral-spatial features in a joint manner. This study presents a performance comparison of recently proposed CNN models, namely 1D CNN, 2D CNN, 3D CNN, and recently introduced fusion based CNNs has been presented. Three HSI benchmark datasets including were used to assess the classification performance.